AI Chatbots Are Steering Voters Toward Bad Election Information

AI chatbots are already shaping how some Americans understand the 2026 election, but the tools are still getting core facts wrong, relying on partisan or low-quality sources, and sometimes refusing to answer basic election questions. Reporting and testing cited by WIRED shows the risk is immediate: voters are using consumer AI for campaign and ballot information today, even as the systems continue to produce unreliable answers.

What the latest testing shows

WIRED’s reporting ties together a set of recent findings that point in the same direction. About 8% of surveyed U.S. adults said they used AI tools ahead of the midterms for candidate or policy information, while nearly half of surveyed adults have used an AI chatbot at some point. That means the audience for these systems is no longer niche, even if only a minority is turning to them specifically for election help.

The problem is what those voters receive. Forum AI found that almost 40% of mainstream chatbot election answers contained at least one factual error. NewsGuard, meanwhile, reported that mainstream chatbots cited so-called “pink slime” sites about half the time on candidate and policy prompts, with ChatGPT citing those sites in more than 70% of searches in its testing. Grok had the lowest citation rate in that test, at 30%.

How are chatbots getting election questions wrong?

They are mixing stale information, hallucinated details and unreliable sourcing. In AI, “hallucination” is the industry term for confident but false output, and the election use case is especially vulnerable because deadlines, ballot rules and candidate information can change quickly.

In the examples cited by WIRED, Google’s AI Mode would not give the 2020 election results. Google Gemini gave the wrong Pennsylvania mail-in ballot deadline. Forum AI said Claude confused Rep. Don Davis with a former school superintendent of the same name. Those failures are not just abstract model quirks; they can send a voter to the wrong deadline, the wrong person or no answer at all.

Why it matters

This matters because election information is one of the highest-stakes categories of consumer AI use. A mistaken recipe or travel suggestion may be annoying; a wrong ballot deadline or a misleading candidate summary can affect whether someone votes at all, or how they vote.

The danger is not limited to fabricated facts. A chatbot that leans on “pink slime” sites can appear confident while laundering partisan messaging through sources that mimic legitimate local or national newsrooms. That can distort the information environment without looking overtly false to a casual user.

For voters, the risk is subtle but real: the interface feels authoritative, the answer arrives quickly, and the user may not know how to verify what they were given. For campaigns and election officials, that creates a communications problem that traditional fact-checking may not solve quickly enough.

By the numbers

Here are the key figures from the reporting and testing cited by WIRED.

Background: why election prompts are hard for AI

Chatbots generate responses using large language models trained on broad web data. That design is useful for general conversation, but it is a poor fit for fast-moving civic information. Election rules vary by state, deadlines shift, and candidate details can be confused when names are similar. The model may also surface whatever material is easiest to retrieve, not what is most trustworthy.

That is where the sources matter. “Pink slime” news sites imitate real outlets but are funded for political or ideological purposes. If an AI system repeatedly pulls from those pages, it can reproduce one-sided or misleading content while keeping the appearance of neutrality. The user may not see a source list, or may not recognize the source as unreliable even if they do.

The WIRED report suggests the issue is not confined to one company. OpenAI, Anthropic and Google have all faced criticism over election-related chatbot accuracy. OpenAI referred WIRED to its 2026 elections blog rather than commenting directly, a sign that the companies know the issue is sensitive even if they are not yet offering a full public fix.

What the companies are saying, and not saying

So far, the public response has been limited. OpenAI did not provide a direct comment to WIRED, instead pointing to its elections blog. The fact sheet does not note direct responses from Anthropic or Google, but their products are part of the criticism because they were among the systems tested or cited in the reporting.

The companies are being evaluated on two fronts at once: whether the systems answer at all, and whether they answer responsibly. Google’s AI Mode refusing to give 2020 election results may look like caution, but Gemini’s wrong Pennsylvania deadline shows that a cautious product can still fail in practice. A filter can block one type of answer while leaving other mistakes intact.

What did researchers actually find?

Forum AI and NewsGuard each tested the tools from a different angle, but both reached troubling conclusions. Forum AI focused on factual accuracy in election answers; NewsGuard focused on the kinds of sources the tools cite. Together, their findings suggest that the reliability problem is both factual and structural.

• Forum AI found nearly 40% of mainstream chatbot election answers had at least one factual error.
• NewsGuard found mainstream chatbots cited “pink slime” sites about half the time on candidate and policy prompts.
• ChatGPT cited those sites in more than 70% of searches in NewsGuard’s testing.
• Grok had the lowest “pink slime” citation rate in the sample, at 30%.

What voters are likely to miss

The biggest risk is not that every answer is wrong. It is that some of the most important answers are wrong in ways that are hard to spot. A ballot deadline can sound credible even when it is not. A candidate summary can be incomplete while still sounding polished. A refusal to answer can push a user to keep prompting until the system produces a more confident but less reliable response.

That creates a trust problem for election administrators and journalists alike. Traditional reporting can correct the record, but a chatbot answer is personalized, immediate and often framed as a direct reply rather than a list of disputed claims. Once a user has acted on it, the correction may be too late.

“I don’t think that most Americans are prepared to navigate an election in the age of consumer AI.”

That warning from Nina Jankowicz, executive director of the American Sunlight Project, captures the scale of the issue. Consumers are being asked to judge source quality in real time, often without seeing where the answer came from or why the model ranked one source ahead of another.

Jankowicz also told WIRED, “It has to ingest this information from somewhere, and often those sources are not very good.” In practice, that means the quality of an election answer is only as strong as the sources feeding the model and the guardrails around retrieval.

Timeline of key moments

The reporting points to a sequence of warnings rather than a single isolated failure.

What happens next

The big questions are still open. It remains unclear how widespread voter use of AI tools is on election day beyond the Pew survey cited in the report, and it is not yet known whether companies will significantly change how their products handle election questions.

Watch for three things: whether OpenAI, Anthropic, Google or xAI tighten source filtering; whether independent testing continues to find similar error rates across platforms; and whether more voters use chatbots for ballot logistics rather than for candidate research. Those are different use cases, and they may fail in different ways.

The broader question is whether consumer AI can become a reliable election helper before election season ends. For now, the evidence in WIRED’s reporting suggests the tools are still too willing to mix confidence with error, and too likely to lean on weak sources when the facts matter most.

Why readers should pay attention now

If you are voting, the safest approach is still to treat chatbot election answers as a starting point, not a final source. Check deadlines with official election offices, compare candidate information against reputable news outlets, and be cautious if a chatbot gives a confident answer without a clear source trail.

The reporting does not argue that AI is useless in elections. It shows something narrower and more urgent: these tools are already in the civic information loop, but they have not yet earned the level of trust that many users may already be giving them.